diff --git a/include/hcc_detail/hip_hcc.h b/include/hcc_detail/hip_hcc.h
index e79ba9e1dc..a3f9c73a13 100644
--- a/include/hcc_detail/hip_hcc.h
+++ b/include/hcc_detail/hip_hcc.h
@@ -466,6 +466,17 @@ inline std::ostream& operator<<(std::ostream& os, const ihipStream_t& s)
     return os;
 }
 
+inline std::ostream & operator<<(std::ostream& os, const dim3& s)
+{
+    os << '{';
+    os << s.x;
+    os << ',';
+    os << s.y;
+    os << ',';
+    os << s.z;
+    os << '}';
+    return os;
+}
 
 //----
 // Internal event structure:
diff --git a/include/hcc_detail/hip_runtime_api.h b/include/hcc_detail/hip_runtime_api.h
index daf737563e..1ee7d70ef5 100644
--- a/include/hcc_detail/hip_runtime_api.h
+++ b/include/hcc_detail/hip_runtime_api.h
@@ -121,7 +121,6 @@ typedef struct dim3 {
 } dim3;
 
 
-
 /**
  * Memory copy types
  *
diff --git a/src/hip_device.cpp b/src/hip_device.cpp
index 86e8563272..f16ae715c1 100644
--- a/src/hip_device.cpp
+++ b/src/hip_device.cpp
@@ -61,7 +61,7 @@ hipError_t hipGetDeviceCount(int *count)
  */
 hipError_t hipDeviceSetCacheConfig ( hipFuncCache cacheConfig )
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(cacheConfig);
 
     // Nop, AMD does not support variable cache configs.
 
@@ -75,7 +75,7 @@ hipError_t hipDeviceSetCacheConfig ( hipFuncCache cacheConfig )
  */
 hipError_t hipDeviceGetCacheConfig ( hipFuncCache *cacheConfig )
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(cacheConfig);
 
     *cacheConfig = hipFuncCachePreferNone;
 
@@ -89,7 +89,7 @@ hipError_t hipDeviceGetCacheConfig ( hipFuncCache *cacheConfig )
  */
 hipError_t hipFuncSetCacheConfig ( hipFuncCache cacheConfig )
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(cacheConfig);
 
     // Nop, AMD does not support variable cache configs.
 
@@ -104,7 +104,7 @@ hipError_t hipFuncSetCacheConfig ( hipFuncCache cacheConfig )
  */
 hipError_t hipDeviceSetSharedMemConfig ( hipSharedMemConfig config )
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(config);
 
     // Nop, AMD does not support variable shared mem configs.
 
@@ -119,7 +119,7 @@ hipError_t hipDeviceSetSharedMemConfig ( hipSharedMemConfig config )
  */
 hipError_t hipDeviceGetSharedMemConfig ( hipSharedMemConfig * pConfig )
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(pConfig);
 
     *pConfig = hipSharedMemBankSizeFourByte;
 
@@ -184,7 +184,7 @@ hipError_t hipDeviceReset(void)
  */
 hipError_t hipDeviceGetAttribute(int* pi, hipDeviceAttribute_t attr, int device)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(attr, device);
 
     hipError_t e = hipSuccess;
 
diff --git a/src/hip_event.cpp b/src/hip_event.cpp
index 983b190644..1514fc5868 100644
--- a/src/hip_event.cpp
+++ b/src/hip_event.cpp
@@ -69,7 +69,7 @@ hipError_t hipEventCreate(hipEvent_t* event)
 //---
 hipError_t hipEventRecord(hipEvent_t event, hipStream_t stream)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(event, stream);
 
     ihipEvent_t *eh = event._handle;
     if (eh && eh->_state != hipEventStatusUnitialized)   {
@@ -104,7 +104,7 @@ hipError_t hipEventRecord(hipEvent_t event, hipStream_t stream)
 //---
 hipError_t hipEventDestroy(hipEvent_t event)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(event);
 
     event._handle->_state  = hipEventStatusUnitialized;
 
@@ -119,7 +119,7 @@ hipError_t hipEventDestroy(hipEvent_t event)
 //---
 hipError_t hipEventSynchronize(hipEvent_t event)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(event);
 
     ihipEvent_t *eh = event._handle;
 
@@ -148,7 +148,7 @@ hipError_t hipEventSynchronize(hipEvent_t event)
 //---
 hipError_t hipEventElapsedTime(float *ms, hipEvent_t start, hipEvent_t stop)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(ms, start, stop);
 
     ihipEvent_t *start_eh = start._handle;
     ihipEvent_t *stop_eh = stop._handle;
@@ -193,7 +193,7 @@ hipError_t hipEventElapsedTime(float *ms, hipEvent_t start, hipEvent_t stop)
 //---
 hipError_t hipEventQuery(hipEvent_t event)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(event);
 
     ihipEvent_t *eh = event._handle;
 
diff --git a/src/hip_hcc.cpp b/src/hip_hcc.cpp
index 6334b6004d..5bdbefcdfe 100644
--- a/src/hip_hcc.cpp
+++ b/src/hip_hcc.cpp
@@ -1113,7 +1113,7 @@ void decKernelCnt(){
 // Allows runtime to track some information about the stream.
 hipStream_t ihipPreLaunchKernel(hipStream_t stream, dim3 grid, dim3 block, grid_launch_parm *lp)
 {
-	std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(stream, grid, block, lp);
     stream = ihipSyncAndResolveStream(stream);
 #if USE_GRID_LAUNCH_20 
     lp->grid_dim.x = grid.x;
@@ -1143,7 +1143,7 @@ hipStream_t ihipPreLaunchKernel(hipStream_t stream, dim3 grid, dim3 block, grid_
 }
 hipStream_t ihipPreLaunchKernel(hipStream_t stream, size_t grid, dim3 block, grid_launch_parm *lp)
 {
-	std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(stream, grid, block, lp);
     stream = ihipSyncAndResolveStream(stream);
 #if USE_GRID_LAUNCH_20 
     lp->grid_dim.x = grid;
@@ -1174,7 +1174,7 @@ hipStream_t ihipPreLaunchKernel(hipStream_t stream, size_t grid, dim3 block, gri
 
 hipStream_t ihipPreLaunchKernel(hipStream_t stream, dim3 grid, size_t block, grid_launch_parm *lp)
 {
-	std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(stream, grid, block, lp);
     stream = ihipSyncAndResolveStream(stream);
 #if USE_GRID_LAUNCH_20 
     lp->grid_dim.x = grid.x;
@@ -1205,7 +1205,7 @@ hipStream_t ihipPreLaunchKernel(hipStream_t stream, dim3 grid, size_t block, gri
 
 hipStream_t ihipPreLaunchKernel(hipStream_t stream, size_t grid, size_t block, grid_launch_parm *lp)
 {
-	std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(stream, grid, block, lp);
     stream = ihipSyncAndResolveStream(stream);
 #if USE_GRID_LAUNCH_20 
     lp->grid_dim.x = grid;
@@ -1253,8 +1253,8 @@ void ihipPostLaunchKernel(hipStream_t stream, grid_launch_parm &lp)
 // HIP API Implementation
 //
 // Implementor notes:
-// _ All functions should call ihipInit as first action:
-//    std::call_once(hip_initialized, ihipInit);
+// _ All functions should call HIP_INIT_API as first action:
+//    HIP_INIT_API(<function_arguments>);
 //
 // - ALl functions should use ihipLogStatus to return error code (not return error directly).
 //=================================================================================================
@@ -1629,7 +1629,7 @@ void ihipStream_t::copyAsync(void* dst, const void* src, size_t sizeBytes, unsig
 //---
 hipError_t hipHccGetAccelerator(int deviceId, hc::accelerator *acc)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(deviceId, acc);
 
     ihipDevice_t *d = ihipGetDevice(deviceId);
     hipError_t err;
@@ -1649,7 +1649,7 @@ hipError_t hipHccGetAccelerator(int deviceId, hc::accelerator *acc)
 //---
 hipError_t hipHccGetAcceleratorView(hipStream_t stream, hc::accelerator_view **av)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(stream, av);
 
     if (stream == hipStreamNull ) {
         ihipDevice_t *device = ihipGetTlsDefaultDevice();
diff --git a/src/hip_memory.cpp b/src/hip_memory.cpp
index 95429b26e6..d458256fdf 100644
--- a/src/hip_memory.cpp
+++ b/src/hip_memory.cpp
@@ -36,7 +36,7 @@ THE SOFTWARE.
  */
 hipError_t hipPointerGetAttributes(hipPointerAttribute_t *attributes, void* ptr)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(attributes, ptr);
 
     hipError_t e = hipSuccess;
 
@@ -85,7 +85,7 @@ hipError_t hipPointerGetAttributes(hipPointerAttribute_t *attributes, void* ptr)
  */
 hipError_t hipHostGetDevicePointer(void **devicePointer, void *hostPointer, unsigned flags)
 {
-    std::call_once(hip_initialized, ihipInit);
+    HIP_INIT_API(devicePointer, hostPointer, flags);
 
     hipError_t e = hipSuccess;